Inverse design of chemical perturbations for desired cellular state transitions with GEMGen

Abstract Perturbation biology links interventions to resulting cellular changes, providing a basis for understanding and manipulating biological processes. Recent advances in computational modelling have enabled prediction of cellular responses to diverse perturbations. However, many biomedical applications require going beyond response prediction to identify perturbations capable of inducing desired responses, an inverse design problem that remains largely underexplored compared to the forward prediction task. Here we introduce GEMGen, a large language model-based framework for the inverse design of chemical perturbations conditioned on desired cellular state transitions. GEMGen represents transitions as text-based, rank-ordered differential-gene instructions, generates molecular structures autoregressively, and uses a compound-response scorer to prioritize candidates for experimental testing. Across multiple benchmarks, GEMGen recovered small molecules structurally or mechanistically relevant to ground-truth compounds, and also generalized across perturbation modalities by generating chemical interventions from genetic perturbation signatures, outperforming existing perturbation design approaches. Beyond in silico evaluation, when conditioned on transcriptional signatures related to tissue fibrosis, GEMGen yielded eight novel antifibrotic structures from 17 de novo synthesized candidates. Notably, one lead candidate, GEMGen-010, exhibited strong antifibrotic activity in animal models of liver fibrosis without any structural optimization. Together, these results establish GEMGen as an effective approach extending perturbation modeling toward intervention design, broadening the methodology and application scope of perturbation biology.